Databricks launches always-on Genie for long-running enterprise work

Databricks introduced a second generation of Genie on November 11, presenting the product as an always-on AI coworker for long-running enterprise work rather than one that waits for a single prompt.

According to Databricks, the new version can keep a job moving across schedules and event triggers, track progress, request missing information and seek human approval when required. Users can supervise, steer or stop multiple jobs and choose which actions Genie may take within company-set guardrails.

Databricks said the system can work through Genie One, email, Slack and Microsoft Teams. Its examples include monitoring insurance claims for emerging fraud patterns, investigating underperforming marketing campaigns and preparing employees for meetings. These are company-provided use cases, not independently measured outcomes.

Databricks describes the product’s context layer, Genie Ontology, as connecting structured and unstructured data with company definitions, relationships and organizational knowledge. The company said it can extract context from dashboards, queries and notebooks, while external business-intelligence sources can add more business logic. Teams can mark authoritative definitions in Unity Catalog, which Databricks said OntoRank then prioritizes. Ontology Manager gives administrators and subject-matter experts a place to review and refine that context.

Databricks says Genie Search combines an index of content from applications such as Google Drive, Jira, Confluence, Workday and SharePoint with live retrieval through Model Context Protocol connectors. Databricks said Genie can use the index, MCP or SQL for a task and compose an answer with citations. The company said access controls from connected source systems are mapped through Unity Catalog and enforced before results are returned.

Databricks says Unity Gateway centralizes controls over connectors, MCP tools and actions. Administrators can specify approval requirements, inspect traces, and set workspace or per-user budgets, spending caps and alerts. Databricks said Genie acts with the user’s identity and follows Unity Catalog and connected-system permissions. If more access is needed, Genie can prepare a request, while administrators and data owners retain approval authority. Databricks has separately detailed serverless access and cost controls.

Databricks said thousands of customers use Genie in production and thousands of teams use Genie Agents. The opened sources did not independently substantiate those figures. The company also reported results from a benchmark developed with Turing: Genie passed 91% of runs, compared with 45% and 58% for two unnamed frontier agents, while using fewer than half as many tokens on average and running three to four times faster. Databricks ran each agent five times per task in a fictional energy-company workspace; the comparison agents used a toolbelt that excluded Genie MCP. No independent replication was identified in the opened sources.

Genie and Genie Ontology are generally available, according to Databricks. Genie Search, Ontology Manager and the new always-on Genie are due to enter preview and beta programs, but the company did not provide firm start dates or specific price figures in the launch materials.